Research On Generation Adversarial Network Based On Image Transform Self-supervised Model

The development of deep learning has achieved great results in image generation,image migration,etc.However,nowadays,research and progress are more towards the support of large-scale data.Deep networks and large data sets have proposed more for the acquisition and labeling of sample sets.Claim.Generating a confrontation network can alleviate the impact of insufficient sample data on training.However,its gradients are easy to disappear and explode,modes are collapsed,and the generator and discriminator are not easy to converge,which significantly affects the quality and effect of generation.To find a way to effectually draw image features,improve network capability and generation quality,and solve the difficulty of manually labeling sample labels,this paper proposes a new structure of image generation model,and provides a framework for the following multi-task training.Compared with supervised learning,unsupervised learning does not require manually labeled label information in model training.Self-supervised learning uses pre-tasks to extract its self-supervised information from large-scale unsupervised data for supervised learning and training.This paper uses the geometric transformation of the image as the pretext task.It uses the model to learn the target object's characteristic texture and information at different geometric positions to learn semantic information to predict the image's transformation.This paper's model performs multi-scale transformation of angle rotation and plane flip on the generated image and the input image,then uses the transformed label as an artificial label and uses multiple discriminator structures with shared weights to achieve classification loss and change prediction.In this paper,a generator-discriminator network model based on the residual construction is designed to stabilize loss convergence and reduce overfitting.This paper introduces a spectral normalization layer to limit the network gradient,strengthening the network's training stability.Through multiple experimental tests on multiple data sets,this paper has been compared with other GAN models,and verified that this model is better than other comparable models in terms of image generation effect,FID and IS score indicators,and loss function convergence.The paper compares the influence of self-supervised learning and residual structure on the training results on its basis and proves that the model's structure design in this paper is logical.This paper designs the application research of self-supervised learning based on image transformation in generative adversarial networks and proposes improved methods for unstable training and low-quality generation.This paper's model has been effectively improved on generating quality and training convergence,has made significant progress in the extraction and learning of image features.